code/run_full_factor_groot.sh

#!/usr/bin/env bash
set -uo pipefail
# ─────────────────────────────────────────────────────────────────────────────
# run_full_factor_groot.sh — GR00T N1.7 full-factor sweep.
#
# Cell sampling, instruction format, header and the per-cell / overall result
# format are COPIED VERBATIM from
#   eval_pi0_5/examples/maniskill_full_factor/run_full_factor_inference.sh
# so the produced full_factor_<ckpt>_seed<NN>.txt files are directly comparable
# to the pi0.5 ones.  Only the per-cell python call drives a GR00T zmq server
# (groot_full_factor_main.py) instead of the openpi websocket policy.
#
# A GR00T inference server must already be reachable at ${HOST}:${PORT}.
#
# Usage:
#   bash run_full_factor_groot.sh [total_episodes] [results_txt_path] [sample_n]
#
# Env overrides: HOST PORT SIM_BACKEND RENDER_BACKEND MAX_EPISODE_STEPS
#                SEED_BASE NO_DISTRACTOR_PROB SAMPLE_SEED REPLAN_STEPS
#                MS_PY GROOT_FF_MAIN VIDEO_ROOT MANISKILL_CONFLICT_ROOT
# pi0.5 protocol values: sample_n=200 sample_seed=42 total_episodes=200
#                        max_episode_steps=500 no_distractor_prob=0.70
#                        sim/render=cpu  seed_base ∈ {40,41,42}
# ─────────────────────────────────────────────────────────────────────────────

TOTAL_EPISODES_TARGET="${1:-200}"
RESULTS_TXT_PATH="${2:-}"
SAMPLE_N="${3:-200}"

HOST="${HOST:-127.0.0.1}"
PORT="${PORT:-5555}"
SIM_BACKEND="${SIM_BACKEND:-cpu}"
RENDER_BACKEND="${RENDER_BACKEND:-cpu}"
MAX_EPISODE_STEPS="${MAX_EPISODE_STEPS:-500}"
SEED_BASE="${SEED_BASE:-40}"
NO_DISTRACTOR_PROB="${NO_DISTRACTOR_PROB:-0.70}"
SAMPLE_SEED="${SAMPLE_SEED:-42}"
REPLAN_STEPS="${REPLAN_STEPS:-5}"

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
MS_PY="${MS_PY:-/workspace/groot_eval/.venv_ms/bin/python}"
GROOT_FF_MAIN="${GROOT_FF_MAIN:-${SCRIPT_DIR}/groot_full_factor_main.py}"

if [[ -z "${RESULTS_TXT_PATH}" ]]; then
  ts="$(date +%Y%m%d_%H%M%S)"
  RESULTS_TXT_PATH="/workspace/groot_eval/results_af/full_factor_ep${TOTAL_EPISODES_TARGET}_sample${SAMPLE_N}_${ts}.txt"
fi
mkdir -p "$(dirname "${RESULTS_TXT_PATH}")"
VIDEO_ROOT="${VIDEO_ROOT:-$(dirname "${RESULTS_TXT_PATH}")/videos_seed${SEED_BASE}}"

# ── Generate task list: all 4320 or a random subset of sample_n ──
# (VERBATIM from pi0.5 run_full_factor_inference.sh)
mapfile -t TASK_LINES < <(python3 - "${SAMPLE_N}" "${SAMPLE_SEED}" <<'PY'
import itertools, random, sys

VERBS    = ["lift","grasp","push","pull","rotate","slide"]
COLORS   = ["red","yellow","blue","orange","green","black"]
SHAPES   = ["cube","sphere","cup","car","pyramid","star"]
SPATIALS = ["left","right","middle","front","behind"]
SIZES    = ["small","large","smaller","larger"]

all_tasks = list(itertools.product(VERBS, COLORS, SHAPES, SPATIALS, SIZES))
n    = int(sys.argv[1])
seed = int(sys.argv[2])
if n > 0:
    rng = random.Random(seed)
    rng.shuffle(all_tasks)
    all_tasks = all_tasks[:n]
for t in all_tasks:
    print(" ".join(t))
PY
)

TOTAL_CELLS="${#TASK_LINES[@]}"
NUM_EPISODES="$(python3 -c "import math; print(math.ceil(${TOTAL_EPISODES_TARGET} / ${TOTAL_CELLS}))")"
echo "sample_n=${SAMPLE_N} → ${TOTAL_CELLS} cells, ${NUM_EPISODES} episodes/cell (target ${TOTAL_EPISODES_TARGET} total)"

{
  echo "# Full-factor inference (GR00T N1.7)"
  echo "sample_n=${SAMPLE_N}  sample_seed=${SAMPLE_SEED}  total_cells=${TOTAL_CELLS}"
  echo "total_episodes_target=${TOTAL_EPISODES_TARGET}  num_episodes_per_cell=${NUM_EPISODES}"
  echo "total_episodes_actual=$((TOTAL_CELLS * NUM_EPISODES))"
  echo "host=${HOST}  port=${PORT}"
  echo "sim_backend=${SIM_BACKEND}  render_backend=${RENDER_BACKEND}"
  echo "max_episode_steps=${MAX_EPISODE_STEPS}  seed_base=${SEED_BASE}"
  echo "no_distractor_prob=${NO_DISTRACTOR_PROB}  replan_steps=${REPLAN_STEPS}"
  echo
  echo "index verb color shape spatial size prompt successes/total"
} > "${RESULTS_TXT_PATH}"

# ManiSkill C-extensions link libtorch.so — make ms-venv torch libs findable.
_MS_TORCH_LIB="$("${MS_PY}" -c 'import torch,os;print(os.path.join(os.path.dirname(torch.__file__),"lib"))' 2>/dev/null || true)"
export LD_LIBRARY_PATH="${_MS_TORCH_LIB}:${LD_LIBRARY_PATH:-}"
export MANISKILL_CONFLICT_ROOT="${MANISKILL_CONFLICT_ROOT:-/workspace/groot_eval/genie_repo/maniskill_conflict}"

total_success=0
total_episodes_done=0
i=0

for line in "${TASK_LINES[@]}"; do
  read -r verb color shape spatial size <<<"${line}"
  i=$((i + 1))

  verb_cap="${verb^}"
  case "${spatial}" in
    left)   phrase="on the left"   ;;
    right)  phrase="on the right"  ;;
    middle) phrase="in the middle" ;;
    front)  phrase="in front"      ;;
    behind) phrase="at the back"   ;;
  esac
  prompt="${verb_cap} the ${size} ${color} ${shape} ${phrase}."
  seed=$((SEED_BASE + i))

  echo "[${i}/${TOTAL_CELLS}] ${prompt}"

  run_log="$(mktemp)"
  "${MS_PY}" "${GROOT_FF_MAIN}" \
    --host "${HOST}" \
    --port "${PORT}" \
    --verb "${verb}" \
    --color "${color}" \
    --shape "${shape}" \
    --spatial "${spatial}" \
    --size "${size}" \
    --num-episodes "${NUM_EPISODES}" \
    --max-episode-steps "${MAX_EPISODE_STEPS}" \
    --sim-backend "${SIM_BACKEND}" \
    --render-backend "${RENDER_BACKEND}" \
    --replan-steps "${REPLAN_STEPS}" \
    --no-distractor-prob "${NO_DISTRACTOR_PROB}" \
    --seed "${seed}" \
    --video-out-path "${VIDEO_ROOT}/${verb}_${size}_${color}_${shape}_${spatial}" \
    2>&1 | tee "${run_log}"
  py_status="${PIPESTATUS[0]}"

  if [[ "${py_status}" -ne 0 ]]; then
    echo "${i} ${verb} ${color} ${shape} ${spatial} ${size} ERROR" >> "${RESULTS_TXT_PATH}"
    rm -f "${run_log}"
    echo "Task ${i}/${TOTAL_CELLS} failed. Partial results: ${RESULTS_TXT_PATH}"
    exit "${py_status}"
  fi

  success_info="$(python3 - "${run_log}" <<'PY'
import re, sys
from pathlib import Path
txt = Path(sys.argv[1]).read_text(encoding="utf-8", errors="ignore")
m = re.search(r"Success rate:\s*(\d+)\s*/\s*(\d+)", txt)
print(f"{m.group(1)} {m.group(2)}" if m else "NA NA")
PY
)"
  rm -f "${run_log}"

  read -r ep_succ ep_total <<<"${success_info}"
  if [[ "${ep_succ}" == "NA" ]]; then
    cell_result="NA"
  else
    total_success=$((total_success + ep_succ))
    total_episodes_done=$((total_episodes_done + ep_total))
    cell_result="${ep_succ}/${ep_total}"
  fi

  echo "${i} ${verb} ${color} ${shape} ${spatial} ${size} \"${prompt}\" ${cell_result}" >> "${RESULTS_TXT_PATH}"
done

overall_rate="$(python3 -c "
s, n = ${total_success}, ${total_episodes_done}
print(f'{100.0*s/n:.1f}' if n > 0 else '0.0')
")"

{
  echo
  echo "overall_success=${total_success}/${total_episodes_done} (${overall_rate}%)"
} >> "${RESULTS_TXT_PATH}"

echo ""
echo "Done: ${total_episodes_done} episodes across ${TOTAL_CELLS} cells"
echo "Overall: ${total_success}/${total_episodes_done} (${overall_rate}%)"
echo "Results: ${RESULTS_TXT_PATH}"